services/ner.py

"""
services/absa.py
Aspect-Based Sentiment Analysis (ABSA) untuk Bahasa Indonesia.

Pendekatan:
1. Ekstrak aspek dari teks menggunakan lexicon + dependency pattern
2. Tentukan sentimen per aspek menggunakan window context
3. Agregasi hasil per kategori aspek

Kategori aspek yang didukung (domain-agnostic):
  - harga/biaya       : harga, mahal, murah, biaya, tarif, ongkos
  - kualitas/produk   : kualitas, bagus, jelek, rusak, bagus, produk
  - pelayanan/service : pelayanan, layanan, respon, lambat, cepat, ramah
  - lokasi/tempat     : lokasi, tempat, jarak, strategis, jauh, dekat
  - kebijakan         : kebijakan, aturan, regulasi, keputusan, program
  - pemimpin/tokoh    : pemimpin, presiden, gubernur, menteri, pejabat
  - ekonomi           : ekonomi, inflasi, harga, pendapatan, gaji, subsidi
  - pendidikan        : pendidikan, sekolah, kampus, belajar, kurikulum
  - kesehatan         : kesehatan, rumah sakit, dokter, obat, vaksin
  - infrastruktur     : jalan, infrastruktur, gedung, fasilitas, listrik
"""

import re
from collections import defaultdict
from typing import Optional

# ─────────────────────────────────────────────
# ASPECT LEXICON
# ─────────────────────────────────────────────
ASPECT_LEXICON = {
    'harga': [
        'harga','mahal','murah','biaya','tarif','ongkos','harganya',
        'cost','price','bayar','bayaran','budget','anggaran','tagihan',
        'cicilan','kredit','diskon','promo','gratis','terjangkau'
    ],
    'kualitas': [
        'kualitas','bagus','jelek','buruk','rusak','cacat','produk',
        'barang','mutu','kualiti','quality','performa','fitur','spesifikasi',
        'durable','tahan lama','awet','rapuh','boros'
    ],
    'pelayanan': [
        'pelayanan','layanan','servis','service','respon','respons','lambat',
        'cepat','ramah','kasar','profesional','sopan','membantu','helpful',
        'cs','customer service','admin','operator','staff','petugas'
    ],
    'lokasi': [
        'lokasi','tempat','jarak','strategis','jauh','dekat','akses',
        'parkir','alamat','wilayah','daerah','kawasan','lingkungan'
    ],
    'kebijakan': [
        'kebijakan','aturan','regulasi','keputusan','program','peraturan',
        'undang','hukum','sanksi','denda','izin','prosedur','birokrasi',
        'pemerintah','pemerintahan','politik','implementasi'
    ],
    'pemimpin': [
        'pemimpin','presiden','gubernur','menteri','pejabat','bupati',
        'walikota','anggota','dewan','partai','calon','kandidat','tokoh',
        'figur','kepala','direktur','ceo','pimpinan'
    ],
    'ekonomi': [
        'ekonomi','inflasi','deflasi','pendapatan','gaji','upah','subsidi',
        'pajak','ekspor','impor','investasi','pertumbuhan','resesi','utang',
        'pinjaman','modal','bisnis','usaha','umkm'
    ],
    'pendidikan': [
        'pendidikan','sekolah','kampus','belajar','kurikulum','guru','dosen',
        'mahasiswa','siswa','nilai','ujian','beasiswa','biaya sekolah',
        'spp','kuliah','universitas','sd','smp','sma'
    ],
    'kesehatan': [
        'kesehatan','rumah sakit','dokter','obat','vaksin','rs','puskesmas',
        'bpjs','asuransi','rawat','operasi','penyakit','covid','virus',
        'faskes','apotek','tenaga medis','perawat'
    ],
    'infrastruktur': [
        'jalan','infrastruktur','gedung','fasilitas','listrik','air','banjir',
        'macet','transportasi','tol','jembatan','bandar udara','pelabuhan',
        'internet','sinyal','jaringan','konstruksi'
    ],
}

# ─────────────────────────────────────────────
# SENTIMENT LEXICON PER ASPECT
# ─────────────────────────────────────────────
SENTIMENT_POS = {
    'bagus','baik','bagus','mantap','keren','hebat','suka','senang','puas',
    'meningkat','naik','maju','berkembang','berhasil','sukses','bagus',
    'terjangkau','murah','gratis','ramah','cepat','tepat','profesional',
    'strategis','dekat','mudah','lancar','aman','nyaman','bersih',
    'good','great','nice','excellent','best','amazing','happy','love',
    'wonderful','perfect','outstanding','satisfied','recommended',
    'mendukung','setuju','approve','pro','positif','memuji','bangga',
}

SENTIMENT_NEG = {
    'buruk','jelek','rusak','parah','kecewa','mahal','lambat','lama',
    'susah','sulit','ribet','boros','kasar','curang','korup','gagal',
    'turun','menurun','anjlok','jatuh','krisis','masalah','bermasalah',
    'berbahaya','bahaya','mengecewakan','tidak puas','kapok',
    'bad','worst','terrible','awful','poor','horrible','hate','dislike',
    'expensive','slow','failed','disappointed','useless','waste',
    'menolak','menentang','against','kontra','negatif','mencela','kritik',
    'bohong','tipu','menipu','korupsi','tidak setuju',
}

NEGATION_WORDS = {
    'tidak','bukan','belum','tak','gak','ga','nggak','ngga','jangan',
    'no','not','never','dont',"don't",'without','tanpa',
}

INTENSIFIER_POS = {'sangat','banget','sekali','amat','luar biasa','super','paling','bgt'}
INTENSIFIER_NEG = {'kurang','agak','sedikit','hampir','nyaris'}


def _get_aspect(token: str) -> Optional[str]:
    """Cari aspek untuk satu token."""
    token = token.lower()
    for aspect, keywords in ASPECT_LEXICON.items():
        if token in keywords or any(kw in token for kw in keywords if len(kw) > 4):
            return aspect
    return None


def _sentiment_score_window(tokens: list, center_idx: int, window: int = 4) -> float:
    """
    Hitung skor sentimen dalam window ±N kata dari posisi aspek.
    Pertimbangkan negasi dan intensifier.
    Return: float positif = positif, negatif = negatif, 0 = netral
    """
    start = max(0, center_idx - window)
    end   = min(len(tokens), center_idx + window + 1)
    window_tokens = tokens[start:end]

    score    = 0.0
    negated  = False
    intensify = 1.0

    for i, tok in enumerate(window_tokens):
        tl = tok.lower()
        if tl in NEGATION_WORDS:
            negated = True
            continue
        if tl in INTENSIFIER_POS:
            intensify = 1.5
            continue
        if tl in INTENSIFIER_NEG:
            intensify = 0.6
            continue

        if tl in SENTIMENT_POS:
            s = 1.0 * intensify
            score += -s if negated else s
            negated = False
            intensify = 1.0
        elif tl in SENTIMENT_NEG:
            s = -1.0 * intensify
            score += -s if negated else s
            negated = False
            intensify = 1.0

    return score


def _score_to_label(score: float) -> str:
    if score > 0.3:   return "Positive"
    if score < -0.3:  return "Negative"
    return "Neutral"


def extract_aspects(text: str) -> list[dict]:
    """
    Ekstrak aspek dan sentimen dari satu teks.

    Return: list of {aspect, sentiment, score, mention, context}
    """
    if not text or len(text.strip()) < 5:
        return []

    # Tokenisasi sederhana
    clean  = re.sub(r'[^\w\s]', ' ', text.lower())
    tokens = clean.split()

    results   = []
    seen_aspects = set()

    for i, token in enumerate(tokens):
        aspect = _get_aspect(token)
        if aspect is None:
            continue

        # Hindari duplikat aspek dalam satu kalimat
        if aspect in seen_aspects:
            continue
        seen_aspects.add(aspect)

        score   = _sentiment_score_window(tokens, i)
        label   = _score_to_label(score)

        # Context window untuk display
        start   = max(0, i - 3)
        end     = min(len(tokens), i + 4)
        context = ' '.join(tokens[start:end])

        results.append({
            'aspect':    aspect,
            'sentiment': label,
            'score':     round(score, 3),
            'mention':   token,
            'context':   context,
        })

    return results


def analyze_absa(texts: list[str]) -> dict:
    """
    Jalankan ABSA pada list teks.

    Return:
    {
      'per_text': list of per-text results,
      'aggregate': {aspect: {Positive: N, Negative: N, Neutral: N, dominant: str}},
      'top_aspects': sorted list of most-mentioned aspects,
      'aspect_sentiment_map': {aspect: dominant_sentiment}
    }
    """
    per_text  = []
    aggregate = defaultdict(lambda: {'Positive': 0, 'Negative': 0, 'Neutral': 0, 'total': 0})

    for text in texts[:80]:   # batasi untuk performa
        aspects = extract_aspects(text)
        per_text.append({'text': text[:100], 'aspects': aspects})
        for a in aspects:
            aggregate[a['aspect']][a['sentiment']] += 1
            aggregate[a['aspect']]['total']         += 1

    # Kalkulasi dominan per aspek
    agg_result = {}
    for aspect, counts in aggregate.items():
        t = counts['total'] or 1
        dominant = max(
            ['Positive', 'Negative', 'Neutral'],
            key=lambda s: counts[s]
        )
        agg_result[aspect] = {
            'Positive':   counts['Positive'],
            'Negative':   counts['Negative'],
            'Neutral':    counts['Neutral'],
            'total':      counts['total'],
            'pos_pct':    round(counts['Positive'] / t * 100, 1),
            'neg_pct':    round(counts['Negative'] / t * 100, 1),
            'neu_pct':    round(counts['Neutral']  / t * 100, 1),
            'dominant':   dominant,
        }

    # Sort by total mentions
    top_aspects = sorted(
        agg_result.items(),
        key=lambda x: x[1]['total'],
        reverse=True
    )

    aspect_sentiment_map = {
        asp: data['dominant']
        for asp, data in top_aspects
    }

    return {
        'per_text':            per_text[:20],   # kirim sample ke frontend
        'aggregate':           agg_result,
        'top_aspects':         [{'aspect': a, **d} for a, d in top_aspects[:8]],
        'aspect_sentiment_map': aspect_sentiment_map,
        'total_texts_analyzed': len(texts),
        'aspects_found':       len(agg_result),
    }